Sachin Kansal doesn’t raise his voice. Uber’s chief product officer, in an interview with TechCrunch, lets slip a vision that few catch on first listen: the company isn’t just sprinkling AI here and there; it’s building a proprietary data geography that changes the balance of the entire stack. Three seemingly separate chapters – financial services, the relationship with Waymo for robotaxis, and the new AV Labs division – form a single design where technological sovereignty becomes the most valuable competitive variable.

Start with the most consequential piece: AV Labs. It’s not a generic research center; it’s the data collection and annotation infrastructure for autonomous driving. Uber knows that every mile driven by self-driving vehicles generates streams of information that are more valuable if they remain under direct control, away from third parties and cloud-only logic. Behind that choice is a message for anyone developing models at scale: for geospatial data, user behavior, and mobility information, on-premise or private edge deployment is not a relic of the past, but the only way to avoid strategic dependencies and ensure compliance in markets like the European Union. It’s no coincidence that GDPR has made data residency a hot topic: an in-house AV Labs allows Uber to build an inference and training engine aligned with the toughest rules, without performance compromises.

The second piece is the Waymo relationship. The Alphabet company remains the reference partner for robotaxi technology, but the interview reveals an undeclared tension: Uber wants the driving hardware and software to stay under the parent’s hood, while the data generated during rides – routes, routing choices, traffic conditions – ends up in its own vault. It’s a chess game where the real prize isn’t the LIDAR sensor, but control of distributed inference and ownership of the dataset to improve models in a virtuous loop. For those evaluating hybrid deployments, this pattern shows that it’s no longer enough to calculate TCO by looking only at GPU costs: the discriminating factor is the ability to orchestrate pipelines that mix cloud inference for non-critical low-latency tasks and on-premise nodes for processing that touches sensitive data.

Financial services complete the picture. Uber Money and its credit ambitions reveal risk scoring models that, to respect user privacy, push toward local execution on mobile devices or corporate servers. Here AI begins to show up tangibly for riders and drivers: not just chatbots, but systems that decide spending limits, loan conditions, detect fraud – all functions where sending raw data to the cloud would pose a legal and reputational risk. Uber, without declaring it, is designing an architecture that keeps computation as close to the source as possible, just as banks do with their on-site anti-money laundering models.

Outside observers might think the company is aiming to be “everything for everyone.” Instead, Kansal’s statement is a manifesto of specialization: Uber concentrates resources only on the AI that serves its vertical ecosystem, and every stack component – data, training, inference – is evaluated based on the control it guarantees, not just the immediate cost. It’s a posture that signals, structurally, that the industrial AI game is shifting from “how much you spend in the cloud” to “how sovereign you are over your information assets.” Winners reduce interdependencies at critical moments.

The third-order implications touch the hardware market. If more companies follow the route Uber has traced, demand for specialized on-premise inference chips and reticulated storage solutions will only grow, leapfrogging the narrative that everything had already migrated to the public cloud. The autonomous vehicle sector, with its low-latency and high-reliability needs, is already rewarding hybrid architectures that integrate GPUs and proprietary accelerators in local nodes. And Uber’s lesson is that this choice isn’t just technical – it forges a new pact with regulators and customers.